So they need brand new top-of-the-line GPUs, but they’re also capable of running on old unpatched IoT devices?
Seems kind of impractical, why even bother with trying to propagate a 4b model running at 0.39 tokens per second if you could instead make a conventional worm using a more powerful AI model targeting known devices? It seems like the only reason you would want to do this is if you can only find vulnerabilities while actually on the given network, but that would be kind of niche since how many people are really running totally custom stuff?
Yeah, hard to get local LLMs running decently fast even when done deliberately. Like people aren’t going to notice their CPU, memory, and GPU are all getting pegged when this malware runs?
Come on, Anthropic, you need to do better than this if you want your “open LLMs dangerous” marketing campaign to be credible.
Yeah the actually scary kind of AI worm is one that scrapes dozens to hundreds of credentials people accidentally pushed to GitHub and spreads its API requests across all of them to minimize the chance of being noticed. Not something that requires downloading 5GB+ of LLM to someone’s potato ass hardware and running it at half a token per second.
You are completely missing the point.
This is done exclusively for doomer headlines.
OpenAI: “the AI escaped and hacked websites”
Us: “Ackshually it didn’t it just connected to the internet, it was always on your server where you could shut it down.”
Anthropic: “Hold my beer.”
Edit: 3 months later:
Anthropic: “See? And this, us why we’re the only ones to be trusted with LLMs, and we must ban open weights”
Advancing a favorable narrative does seem like the reason for this report, but knowing that Anthropic is very biased isn’t enough by itself to say their claims are bullshit. It’s still worth considering and critiquing the argument on its own merits, stopped clock is right twice a day and all that.
Even if I took every word of this at face value, I feel like people would notice the CPU usage pretty quickly, from the heat coming off the chip if nothing else. Local inference is about as subtle as a coinminer.
assuming you need to be fast that’s true, but if you don’t care how much time you take then you can spread the processing out… this is exactly how MoE models are much easier to run locally, but you could go much further
I’m sure the answer that comes out of your throttled CPU only language model in a month is still going to be relevant.
Thankfully, the huggingface incident proved how much care AI companies put in making sure their models cannot break out, and how aligned to human values the LLM swarms are.\s With that said I think regular personal computers are more likely to get hit by this. About as secure (not very) but much more GPU power and data to harvest.
The AI companies will argue that we should therefore not be allowed to own computers powerful enough to run local LLMs, and instead should own thin clients and subscribe.
Not that relevant. Those fears are about malicious use of open weights models, this is about autonomous self replication of (small) models. Clearly the article by the AI company describes this potentially happening on computers with tiny memory and processing power, where even thin clients could be effected. I also disagree strongly that that’s what the AI companies want, especially the employees which I have never seen espouse anything like that. They just don’t want some random teenager creating the next covid. There are ways to do that without meaningfully limiting compute for individuals.
